A Slimmer and Deeper Approach to Network Structures for Image Denoising and Dehazing

Image denoising and dehazing are representatives of low-level vision tasks. The right trade-off between depth and computational complexity of convolutional neural networks (CNNs) is of significant importance to these problems. Wider feature maps and deeper network are beneficial for better performance, but would increase their complexity. In this paper, we explore a new way in network design, to encourage more convolution layers while decrease the width of feature maps. Such slimmer and deeper architectures can enhance the performance while maintain the same level of computational costs. We experimentally evaluate the performances of the proposed approach on denoising and dehazing, and the results demonstrate that it can achieve the state-of-the-art results on both quantitative measures and qualitative performances. Further experiments also indicate that the proposed approach can be adapted for other image restoration tasks such as super-resolution.